Object Tracking Paper (3): SP-KCF---Deformable Patches using Shape-Preserved KCF

Time: January 30th, 2018, the twentieth week

Weekly Summary: In this week, I continued to read a paper about CF. And this paper is a good improvement of KCF deploying deformable patches.

Paper: Non-Rigid Object Tracking via Deformable Patches using Shape-Preserved KCF and Level Sets/ Author: Xin Sun, Ngai-Man Cheung, Hongxun Yao, Yiluan Guo/ Publication information: ICCV 2017

Outline: This algorithm is the improvement of KCF. It divides the target into deformable patches, and use level sets to efficiently represent the contours and patches.  Based on the current observation, it will track the individual path independently. In this process, the patches are tracked by kernelized correlation filter. When the curve is updated, the tracker will give each one a confidence score, and seek the max probability curves. And they design an evaluation process to distinguish the discriminative patches, employing the photometric discrimination and shape variation. And finally coordinate the good patches to obtain a collaboration solution.

Advantages: This method improve the traditional KCF tracker very much. And it's very accurate and robust when tracking non-rigid object. Those patches can be deformable according their contour. It is not solid or rectangle.

Disadvantages: It is very disappointing that the patches in the first frame is manually labeled. And the number os the patches will decrease the speed of the algorithm.